Emotional Information Features And Its Classification Based On EEG Signals

Emotion is an important part of human psychology.It is a foundation of social communication to recognize the emotion correctly and then make a reasonable response.With the development of brain-computer interfaces(BCI)technology,the human emotional information is introduced into the BCI system so as to make the process of human-computer interaction natural and real.Therefore,researches on emotion recognition based on electroencephalography(EEG)signals,which permits computer to directly recognize a user's emotional state and make humanized responses,is currently carried out.As one of the most common psychological problems,depression brings huge negative effects to people's work and life.It is of great importance to recognize the depressive emotion timely and accurately and then make appropriate adjustment.On the other hand,as a main way of human self-expression in daily life,language is closely related to human emotional expression.Nonetheless,compared with other emotional materials such as emotional pictures and faces,language evokes a relatively low arousal emotion.Hence,language is rarely used as stimulus materials in the existing researches of emotional EEG recognition.Moreover,the most researches of language emotion focus on event related potential(ERP),while emotion recognition based on single-trial EEG signals could be particularly important for BCI systems.Therefore,the author designs two experiments of emotion processing in Chinese language,and performs classification of the emotional information of the language materials and the depressive subjects by using single-trial EEG signals.The main work includes the following four parts:1)Experimental design and data acquisition of emotion processing in Chinese language.Experiment one:Chinese two-character words experiment.Experimental material consists of emotional positive,negative and neutral Chinese two-character words,and unreal two-character words.All materials are presented individually and randomly,and the subjects are asked to press a key for the unreal words.Experiment two:Chinese sentences experiment.Each neutral word is the ending word of a pair of sentences,one neutral sentence and one emotional sentence,which are pair-matched.Both types of sentences' contexts are emotionally neutral.However,for the emotional sentences,when the final words are integrated with the contexts,the events described by the sentences are considered to be negative.Sentences are presented word by word.Subjects are asked to judge whether a probe word following a sentence is in the sentence that they have just read,by pressing two different key-presses.2)The recognition of emotional information of Chinese words.(a)Based on EEG signals evoked by emotional two-character words,emotional features are extracted by common spatial patterns(CSP),and then are classified by using linear discriminant analysis(LDA).(b)Whereas the traditional CSP is disadvantageous on small size training sample data,the temporally regularized CSP(TRCSP)algorithm is proposed.TRCSP obtains the temporal structure information of EEG signal using locally linear embedding,which is embedded into the objective function of CSP as a penalty term.After being verified effective on the public data sets from BCI competitions,TRCSP is applied to the problem of emotion recognition.The results of permutation tests confirm the statistical significance of the recognition accuracy,and reveal the identifiability of the emotional information of Chinese words based on EEG signals.The classification results show that:The recognition rate of negative words versus neutral words is higher than that of positive words versus neutral words for the normal control group,while opposite in the depression group.It demonstrates that the depressive subjects have different emotional processing mechanism from the normal control subjects.The results also indicate the subjects with depressive emotion can be identified.3)The recognition of emotional information of Chinese sentences.The TRCSP algorithms is employed to extract features on the data of emotional sentences,and then LDA is utilized to classify the features.Permutation tests of the experiment results show that classification results are statistically significant for 9 out of 14 subjects.It confirms the identifiability of the emotional information of Chinese sentences based on EEG signals.4)The recognition of subjects with depressive emotion.Considering the individual differences between subjects,a new formulation of regularized CSP based on transfer learning with weighted subjects(RCSPTLw)is proposed.The RCSPTLw approach transfers the information of existing subjects to testing subjects,maximizing the difference between subjects of different classes while minimizing the difference between subjects of the same classes.After being verified effective on public data sets from BCI competitions,RCSPTLw is applied to recognize subjects with depressive emotion on the data of the words experiment.EEG signals of source subjects are used as training samples while target subjects as testing samples.The experiment results indicate that the depressive subjects can be recognized effectively.5)Furthermore,discriminative spatial patterns(DSP)with transfer learning is likewise applied in the three recognition work above.Compares to CSP,the results of DSP is higher in almost all cases.Considering the individual differences between subjects,a new formulation of regularized DSP with transfer learning(RDSPTL)is also proposed.Great recognition rates further show that the depressive emotion of subjects can be identified.The author designs the experiments of processing of two classes Chinese materials,verifies the identifiability of the emotional information in Chinese language,investigates the recognition of subjects with depressive emotion,and develops the emotion research of BCI system based on EEG signals.Moreover,three new algorithms(TRCSP,RCSPTLw,RDSPTL)are propose,which provide new computing methods for a BCI system.